Defect inspection in semiconductor images using statistical methods and neural networks

Abstract

Most defect inspection methods in semiconductor systems need design layout or golden die image. Unlike most methods that require additional information, this paper presents an automatic inspection of defects in semiconductor images with single image. We devise two method to classify the semiconductor images. The first method is based on the cosine similarity and it classify an image into four types: flat, linear, pattern, and complex. The second method is based on the singular value decomposition and it determine whether the image has a linear or complex structure. For the linear image, we present two method to reconstruct a defect-free image. For the patterned image, we also reconstruct a defect-free image. By subtracting the defect-free image from the input image, we obtain the flat image. Defects are found by estimating the inlier distribution for flat images. The Fast-MCD method estimates the inlier distribution by choosing the subset which has the minimum covariance determinant. We devise a statistical method to estimate the inlier distribution by fitting a log-histogram at the center. An image with a complex structure is inspected by two neural networks. In addition, if computer aided design (CAD) data is available, we use it to construct a signed distance function (SDF). Along the level curve of the SDF, we obtain a defect-free image. Then, flatten image can be obtain from difference image.